What AI Can and Cannot Do

Published ranges from major labour research suggest that between 25% and 40% of current work tasks could be automated using AI at present capability levels (McKinsey, 2023; Goldman Sachs, 2023). That range matters more than the headline: the difference between the low and high end is a different plan of action for your career. This module explains what drives the uncertainty, where the lines actually sit today, and how to read capability claims usefully without panic or hype.

Reading the headline numbers honestly

When you see "40% of jobs will be impacted by AI" in a headline, it is almost certainly a translation of a more careful underlying claim. McKinsey Global Institute estimated in 2023 that generative AI could automate work activities that absorb 60-70% of employee time -- but the same report distinguished between automating tasks and replacing jobs. Goldman Sachs separately estimated that 18% of global work could be automated by AI, with major variation by occupation and country. Stanford HAI tracks year-on-year shifts in what AI systems can actually do on benchmarks; the 2024 AI Index Report noted that AI surpassed humans on several reading and image classification benchmarks while still trailing badly on planning and multi-step reasoning.

None of these is a prediction that a specific role disappears. They are estimates of task-level exposure aggregated upward. Your job is a bundle of tasks, and the exposure of each task is what matters for your planning.

Where AI performs reliably today

Current AI systems -- large language models, vision models, speech-to-text -- are reliable for a specific class of task: pattern completion across language, image, and structured data. In practice this covers summarising documents, drafting structured text from a brief, classifying inputs into predefined categories, extracting structured data from unstructured text, translating between languages, transcribing audio, and generating code from a specification.

The common thread across reliable tasks is two-fold: there is a large body of prior examples the model can learn from, and the output can be verified by a human quickly. When both conditions hold, AI performance is high and the cost of any individual error is small because the human catches it in the verification step.

Where AI struggles

AI systems fail in predictable ways. They struggle with tasks that require maintaining a coherent plan over many steps without explicit tracking, genuine novelty (problems outside their training distribution), sensorimotor work that requires physical feedback, ethical judgment in unfamiliar situations, real-time information not present in training data, and interpersonal dynamics -- reading a room, negotiating under pressure, building long-term trust through repeated small interactions.

These are not temporary gaps that close next quarter. Some reflect hard constraints in current architectures (the model has no body, no continuous memory, no stake in your meeting). Others may close, but on a multi-year timeline. The right move is to plan around the capabilities that exist now and the failure modes that will still exist 18 months from now, not around the forecasts.

Why capability boundaries move -- and what to watch

AI capability is not static. Model releases between 2023 and 2025 expanded reliable performance on coding, multimodal tasks, and long-document reasoning. But progress is uneven. Some tasks that looked close to automation (complex legal analysis, medical diagnosis, contract negotiation) remain heavily human-dependent because error cost is high and edge cases are structurally important. Others jumped faster than most forecasts expected, particularly anything involving code, translation, and structured extraction.

The most useful signal to watch is not "AI is getting smarter" in general but specific task-level claims from published evaluations: the system cards published with each model release from Anthropic, OpenAI, and Google; the Stanford HAI AI Index Report (annual); and OECD occupation-level analysis. These cite what changed, with evidence and benchmarks. Industry forecasts without specific task grounding -- "30% of work will be automated by 2030" without saying which tasks -- are less useful for planning your week, your year, or your next role.

A reliable filter: when a capability claim is made, ask "on what specific evaluation, scored how, against what baseline?" If the answer is "it impressed people in a demo", treat it as a leading indicator and nothing more. If the answer is "GPQA Diamond, scored 84%, human PhD baseline is 65%", you can plan around it.

How to use this in your role

The practical question is not "will AI replace my job" but "which of my tasks fit the pattern-completion profile, and which require judgment, novelty, or interpersonal context?" The next module walks through a structured way to answer that for your specific role. Before you get there, the practice block below will get you started on a working map of your own week.